Output Fluctuations in the United States:

What Has Changed Since the Early 1980s?

Margaret M. McConnell and Gabriel Perez Quiros1

February 10, 2000

1Federal Reserve Bank of New York. The authors thank Paul Bennett, Jordi Gal¶³, JimHamilton, Jim Kahn, Kenneth Kuttner, Jose Lopez, Patricia Mosser, Christopher Sims,Charles Steindel, Kei-Mu Yi, Egon Zakraj·sek, Victor Zarnowitz for helpful comments andsuggestions. Justin McCrary provided excellent research assistance. The views expressedin this paper are those of the authors and do not necessarily re°ect the views of the FederalReserve Bank of New York or the Federal Reserve System. Please direct all correspon-dence to the authors at: meg.mcconnell@ny.frb.org or Federal Reserve Bank of New York,Domestic Research, 33 Liberty Street, New York, NY 10045.

Abstract

We document a structural decline in the volatility of real U.S. GDP growth in the ¯rstquarter of 1984. As a means of understanding the dramatic volatility reduction, wedecompose output growth by major product type and provide evidence that the breakemanates from a reduction in the volatility of durable goods production. We furthershow that the break in durables is roughly coincident with a break in the proportionof durables accounted for by inventories. We note that the break in output volatilitya®ects the implementation of a wide range of simulation and econometric techniquesand o®er one important illustration of this in the context of a regime-switching modelof output growth.

1 Introduction

The business press is currently sprinkled with references to the `death' or `taming'

of the business cycle in the United States. While such claims are at best premature,

they are in part rooted in the apparent reduction in the volatility of U.S. output

°uctuations over the period starting in the early 1980's. Figure 1 plots the growth of

U.S. real GDP over the period 1953:2 to 1999:2; the variance of output °uctuations

over the period ending in 1983 is more than four times as large as the variance for

the period since 1984.

In this paper, we document a structural break in the volatility of U.S. GDP growth

in the ¯rst quarter of 1984. As a means of understanding this dramatic reduction in

volatility, we decompose output growth by major product type and provide evidence

that the aggregate volatility break emanates from a reduction in the volatility of

durable goods production. We show that the break in durables is roughly coincident

with a break in the proportion of durables output accounted for by inventories.

The break in output volatility a®ects the implementation of a range of simulation

and econometric techniques. For example, one common method for taking theory to

the data is to compare the moments of data generated from calibrated models with

the moments of actual data. The presence of a one-time reduction in output volatility

in the early 1980's clearly a®ects the time horizon over which the second and higher

moments of output growth should be computed.

On the empirical front, the volatility break implies that linear models for output

growth over periods that span the break are misspeci¯ed. In addition, signal-to-

noise ratios in state-space characterizations of business cycle °uctuations, such as

dynamic factor or Markov-switching models, will be reduced when the variance is

modeled as constant. We present one important example of this in the paper. Finally,

the reduction in the variance of output °uctuations should alter the interpretation

policymakers place on a particular realization of quarterly GDP growth; what may

have been considered a moderate °uctuation in activity prior to the break may now

be viewed as severe.

1

Figure 1: U.S. Real GDP Growth: 1953:2 to 1999:2

1953.2 1957.2 1961.2 1965.2 1969.2 1973.2 1977.2 1981.2 1985.2 1989.2 1993.2 1997.2-4

-2

0

2

4

The paper proceeds as follows. In Section 2, we use a series of structural stability

tests to characterize the nature and timing of the changes in the process for output in

recent years. We also use a regime-switching framework to examine the implications

of lower quarter-to-quarter output volatility for modeling business cycle °uctuations.

In Section 3, we examine disaggregate output data in order to isolate the source of

the volatility decline. Section 4 concludes.

The paper proceeds as follows. In Section 2, we use a series of structural stability

tests to characterize the nature and timing of the changes in the process for output in

recent years. We also use a regime-switching framework to examine the implications

of lower quarter-to-quarter output volatility for modeling business cycle °uctuations.

In Section 3, we examine disaggregate output data in order to isolate the source of

the volatility decline. We then discuss a set of potential explanations and look at the

extent to which each explanation has its main implications borne out in the data.

Section 4 concludes.

2

2 The Decline in U.S. Output Volatility

There is a large literature exploring the question of whether the magnitude and du-

ration of economic °uctuations have changed across the pre- and post-WWII periods

(examples include DeLong and Summers (1986), Romer (1986a, 1986b, 1989, 1994),

Shapiro (1988), Diebold and Rudebusch (1992), Lebergott (1986) andWatson (1994)).

While the evidence on this issue is mixed (resulting in no small part from the dif-

¯culties associated with the construction of comparable data series across the two

periods), the general pursuit of documenting changes in the process governing output

°uctuations is an important element of macroeconomic research. Such documentation

is valuable for the reasons that it leads to a collection of stylized facts and may also

provide insight into whether such changes are likely to be permanent or temporary.

In this section we characterize recent changes in the process for U.S. output

growth. We focus on quarter-to-quarter °uctuations in the growth rate of GDP,

rather than on changes in the business cycle per se. In addition, since we are inter-

ested in the rather dramatic reduction in output volatility in the most recent two

decades relative to the previous three, we use only post-war data and thereby avoid

the problems associated with pre-and post-war data comparability.

2.1 Structural Change

To complement the ocular evidence presented in Figure 1, we begin our empirical

analysis with a simple nonparametric characterization of the changes in process for

U.S. real GDP over the post-war period. Throughout the paper we use chain-weighted

GDP data, as constructed by the Bureau of Economic Analysis (BEA), for the sample

1953:2 to 1999:2. Looking ¯rst for evidence of instability in the mean growth rate, we

¯t a constant and a linear trend to real GDP growth. This regression yields a negative,

but statistically insigni¯cant, coe±cient on the trend term. The insigni¯cance of the

trend term is robust to the use of the ¯rst di®erence of GDP growth rather than the

level, as well as to the inclusion of a lagged dependent variable.

3

Whatever instability the trend term fails to detect in the mean rate of output

growth, it uncovers easily in the variance. For example, regressing the square of GDP

growth on a constant and a trend yields a coe±cient on the trend that is negative

and signi¯cant at the one percent level. The sign and signi¯cance of the trend term

is robust to the substitution of the absolute value of GDP growth for the square and

to the inclusion of lagged values of the dependent variable. Finally the trend term is

negative and highly signi¯cant when included in the variance term of an ARCH or

GARCH speci¯cation for the residual from an AR(1) speci¯cation for GDP growth.

As an alternative exercise, we remove the mean of GDP growth and perform a

CUSUM and CUSUM of squares test on the residuals. The CUSUM test (Brown,

Durbin, and Evans, 1975) is based on the cumulative sum of the one-step ahead fore-

cast error resulting from a recursive estimation. Instability in the parameters of the

mean is indicated if the cumulative sum goes outside the area between the two critical

lines. The CUSUM of squares test is based on the cumulative sum of the squared

one-step ahead forecast error resulting from a recursive estimation. Movement out-

side the critical lines is suggestive of variance instability. Figure 2 plots these two test

statistics. The CUSUM test, shown in the left panel, does not reveal instability in the

mean, while the CUSUM of squares test, shown in the right panel, detects instability

in the variance beginning in the early 1980s.

Imposing a bit more structure on the problem, we follow Hess and Iwata (1997)

and model GDP growth as an AR(1).1 To detect parameter instability in this model

we use Nyblom's L test as described in Hansen (1992a). Large values of the test

statistic imply a rejection of the null hypothesis of stability. Using the asymptotic

critical values for a 5 percent test from Hansen, we see in Table 1 that we cannot

reject the stability of both the constant term and autoregressive coe±cient over our

sample, but we can reject stability of the variance.2 Finally, we reject the hypothesis

1Hess and Iwata show that a ARIMA(1,1,0) model is at least as good as many widely usednonlinear models at replicating the duration and amplitude of °uctuations in the log of real GDP.In addition we use standard lag selection to criteria to test for the best model for our sample and¯nd that we cannot reject an AR(1) speci¯cation in favor of one with longer lags.

2Hansen (1992a) points out that if both the autoregressive and error variance have shifted, the

4

Figure 2: U.S. Real GDP: 1953:2 to 1999:2

-40

-20

0

20

40

55 60 65 70 75 80 85 90 95

CUSUM 5% Significance

-0.2

0.0

0.2

0.4

0.6

0.8

1.0

1.2

55 60 65 70 75 80 85 90 95

CUSUM of Squares 5% Significance

Table 1: Nyblom's L Test for Stability of U.S. Real GDP Growth - 1953:2 to 1999:2

Speci¯cation: ¢yt = ¹+ Á¢yt¡1 + ²tEstimate Lc CV(5 percent)

¹ 0.50 (0.10) 0.09 0.48Á 0.34 (0.08) 0.10 0.48¾2 0.89 (0.12) 0.82 0.48

Joint Lc 1.06 1.02

Note: Nyblom's L test as described in Hansen (1992). ¢y is real GDP growth. p-values in

parentheses to the right of estimates. Lc is the test statistic for a breakpoint in each of the

coe±cients listed in the ¯rst column. CV(5 percent) is the ¯ve percent critical value for the

null hypothesis of no break.

5

of the joint stability of the parameters at the 2.5 percent level.

2.2 The Timing of the Break

In this section we estimate the timing of the structural change in the process for GDP

growth and assess its statistical signi¯cance. Continuing to model GDP growth as

an AR(1), we test for a structural break in the residual variance from the following

speci¯cation for GDP growth:

¢yt = ¹+ Á¢yt¡1 + ²t (1)

If ²t follows a normal distribution,q¼2j²̂tj is an unbiased estimator of the standard

deviation of ²t. Therefore, we look for a break in an equation of the form:

r¼

2j²̂tj = ®+ ¹t (2)

where ® is the estimator of the standard deviation.3

We estimate a break point by jointly estimating the following system using GMM:

¢yt = ¹+ Á¢yt¡1 + ²t (3)

r¼

2j²̂tj = ®1D1t + ®2D2t + ¹t (4)

where

D1t =

8><>: 0 if t · T1 if t > T

power of the L test to detect the shift in the autoregressive parameter is low and thus this test doesnot allow us to rule out instability in the autoregressive parameter with any degree of certainty.This issue also arises with the tests for structural change used in the next section. At that point weapply a correction suggested in Hansen (1999).

3Our choice of speci¯cation here follows the convention of the ¯nance literature and appeals tothe fact that the absolute value speci¯cation is more robust to departures from conditional normality(as pointed out in Davidian and Carroll (1987)) than the estimator of the variance, ²̂2t .

6

D2t =

8><>: 1 if t · T0 if t > T

and T is the estimated break point, and ®1 and ®2 are the corresponding estimators

of the standard deviation. The instruments for each period t are a constant, ¢yt¡1 ,

D1t, and D2t.

The appearance of the parameter T under the alternative hypothesis but not

under the null implies that the LM, LR and Wald tests of equality of the coe±cients

®1 and ®2 do not have standard asymptotic properties. Andrews (1993) and Andrews

and Ploberger (1994) develop tests for cases such as this, when a nuisance parameter

is present under the alternative but not under the null. They consider the function,

Fn(T ), where n is the number of observations, de¯ned as the Wald or LM statistic

of the hypothesis that ®1 = ®2, for each possible value of T. We assume that T lies

in a range T1; T2, where T1 = :15 ¤ n and T2 = :85 ¤ n. Andrews (1993) shows theasymptotic properties of the statistic:

supT1·T·T2

Fn = supFn(T ) (5)

and reports the asymptotic critical values. In this test, the T that maximizes Fn(T )

will be the estimated date of the break point. Andrews and Ploberger (1994) propose

two additional test statistics:

expFn = ln(1=(T2 ¡ T1 + 1)) ¤T2X

T=T1

exp(1=2 ¤ Fn(T )) (6)

and

ave Fn = (1=(T2¡T1 + 1)) ¤T2X

T=T1

Fn(T ): (7)

The p-values associated with these statistics are computed using the approximation

suggested by Hansen (1997).

7

Table 2: Structural Break Tests: U.S. Real GDP Growth - 1953:2 to 1999:2

Panel ASpeci¯cation: ¢yt = ¹+ Á¢yt¡1 + ²t, ²t » N(0; ¾2t )where ¾2t = ¾

21 if t · T , and ¾2t = ¾22 if t > T

Null Sup Exp Ave¾21 = ¾

22 17.80 6.54 6.71

(0.00) (0.00) (0.00)

Estimated break date: 1984:1

Panel BSpeci¯cation: ¢yt = ¹+ Á¢yt¡1 + ²t, ²t » N(0; ¾2t )

where (¹t = ¹1, Át = Á1) if t · T , and (¹t = ¹2, Át = Á2) if t > TNull Sup Exp Ave Chow

¹1 = ¹2, Á1 = Á2 6.34 0.77 1.11 0.23(0.55) (0.75) (0.78) (0.89)

Estimated break date: none

¹1 = ¹2 1.48 0.24 0.45 0.13(0.99) (0.82) (0.76) (0.71)

Estimated break date:none

Á1 = Á2 2.53 0.34 0.59 0.00(0.85) (0.67) (0.62) (0.99)

Estimated break date: none

Note: Panel A presents the results of structural break tests of the type discussed in Andrews

(1992) and Andrews and Ploberger (1994) for the null hypothesis of no break in the variance

parameter. `Sup', `Exp' and `Ave' refer to the supremum, exponential and average test

statistics described in Equations 5, 6, and 7, respectively. p-values appear in parentheses

below test statistics. Panel B presents the results from same tests for the null hypothesis

of no break in all and each of the parameters for the mean. This panel also presents the

results of a Chow test that imposes the 1984:1 estimated break for all and each of the mean

parameters.

8

The results of the tests for structural change in the residual variance of the process

for the growth rate of GDP are reported in the top panel of Table 2. Each of the

three test statistics indicate a strong rejection of the null hypothesis that ¾1 = ¾2,

and the estimated break date occurs in the ¯rst quarter of 1984.

We next consider the possibility that the observed break in the residual variance

is actually the result of a break in the parameters characterizing the mean of the

output process. We therefore estimate:

¢yt = ¹1D1 + ¹2D2 + Á1¢yt¡1D1 + Á2¢yt¡1D2 + ²t (8)

where D1 and D2 are as de¯ned above, ¯rst testing jointly for a break in the constant

and the AR coe±cient, and then for a break in each separately.

These results are reported in the bottom panel of Table 2. In all cases we cannot

reject the null of no break and therefore conclude that the variance break is not

attributable to a change in the constant and AR component of the model.4 Even

when we conduct Chow tests and impose the estimated break date of 1984:1, we still

cannot reject the null of no break. The p-values and test statistics for the Chow test

are reported in the last column of Table 2.

Finally, one might hypothesize that the break in 1984:1 simply re°ects a return

to stability after the highly volatile 1970s. To investigate this possibility, we test for

additional breaks in the subsamples 1953:2 to 1983:4 and 1984:1 to 1999:2, conditional

on having found the ¯rst break in 1984:1. Though we do not report the results here,

4The p-values presented in Table 2 are based on asymptotic critical values. In order to calculatethe small sample properties of our tests we use a bootstrap technique proposed in Hansen (1999),who points out that the use of standard bootstrap techniques is inappropriate in the presence ofstructural change in the regressors. The p-values obtained using this correction are nearly identicalto those presented here. We also examine the small sample properties of our variance break testsusing the stationary bootstrap proposed by Politis and Romano (1994). This bootstrap methodologyis applicable under the null hypothesis of no break in the variance, given that we reject a break inthe mean using Hansen (1999). For 1000 iterations, we obtain a p-value of 0.004 for the supreme,0.002 for the exponential and 0.004 for the average. Thus we can still strongly reject the hypothesisof no break in the variance.

9

we ¯nd that we cannot reject the hypothesis of no break in either subsample.5

2.3 Implications for Business Cycle Modeling

Hamilton (1989) uses a regime switching framework to show that by allowing the

mean rate of GDP growth to switch between two states, one can capture the periodic

shifts between positive and negative real GDP growth in the U.S. He further shows

that such shifts accord well with the NBER business cycle peaks and troughs. A

number of researchers have since found this to be a useful approach to characterizing

business cycles, including Lam (1990), Boldin (1994), Durland and McCurdy (1994),

Filardo (1994), Kim (1994), and Diebold and Rudebusch (1996).

A typical formulation of this model for GDP growth is one in which the mean of

an autoregressive speci¯cation is allowed to vary across states of the economy (usually

two) and the residual variance is assumed constant. In this section, we augment this

standard formulation in two ways. First, we allow the mean and the variance to

follow independent switching processes. Second, we allow the two-state process for

the mean, which we consider the business cycle component of the model, to vary

according to the state of the variance.

In particular, we estimate the following augmented Markov model for GDP growth:

¢yt = ¹St;Vt + Á(¢yt¡1 ¡ ¹St¡1;Vt¡1) + ²t; ²t » N(0; ¾Vt) (9)

where St and Vt are latent variables for the states of the mean and variance, respec-

tively, of output growth, each of which can independently assume a value of 1 or 2.6

This speci¯cation yields two possible states for the variance, with ¾1 being the vari-

5In independent research Koop and Potter (2000) use a Bayesian methodology to search over anumber of alternative models of output data for the U.S. and U.K. Their results suggest that forU.S. real GDP growth a model with a structural break in the variance in the early- to mid-1980's isthe preferred model.

6Our use of just one autoregressive lag di®ers from Hamilton's AR(4) speci¯cation. For bothHamilton's original constant variance speci¯cation and our more general model, however, an AR(1)speci¯cation ¯t best over our sample.

10

ance in the high variance state and ¾2 the variance in the low variance state, along

with four possible states for the mean of output growth:

¹11 for St = 1; Vt = 1

¹21 for St = 2; Vt = 1

¹12 for St = 1; Vt = 2

¹22 for St = 2; Vt = 2

where ¹11, for example, indicates the mean rate of growth in the high mean, high

variance state and ¹21 is growth in the low mean, high variance state.

The results of this estimation are reported in Table 3. The columns in the top

panel of the table indicate the state of the variance parameter. Note that the residual

variance drops from 0.87 per quarter in the high variance state to 0.14 in the low

variance state.

To aid in the interpretation of the coe±cients reported in the table, the top panel

of Figure 3 plots the smoothed probabilities of the low variance state (V = 2). Con-

sistent with the structural break results presented in the previous section, we ¯nd a

sharp increase in the probability of the low variance state during the years 1983-1984.

Note also that the pattern of smoothed probabilities re°ects a one-time switch to the

low variance state in the early 1980's, rather than a switch out of the low variance

state in the 1970's and a switch back into that state in the early 1980's. The pattern

observed here reinforces our earlier ¯nding that the break in 1984:1 is not simply due

to a return to stability after the highly volatile 1970's.

One powerful feature of the regime-switching model used by Hamilton is its ability

to generate smoothed probabilities that correspond closely to the NBER business

cycle peaks and troughs. However, Hamilton estimated his model using data through

1984, the year in which we date the decline in volatility. Does the addition of the

lower volatility years a®ect the ability of the widely used switching mean, constant

variance speci¯cation to identify periods of recession and expansion?

11

Table 3: Markov-Switching Model: U.S. Real GDP Growth, 1953:2 - 1999:2

V=1 V=2¹1v 1.07 (0.15) 0.76 (0.06)¹2v -0.68 (0.48) -0.54 (0.31)¾2v 0.87 (0.14) 0.14 (0.03)Á 0.12 (0.10)p¹11 0.95 (0.02)p¹22 0.65 (0.16)

p¾2

11 0.99 (0.00)

p¾2

22 0.99 (0.01)Ho : ¹11 = ¹12; ¹21 = ¹22 p-value=0.23

Note: The parameter estimates reported in the table refer to the following model for GDP

growth: ¢yt = ¹St;Vt + Á(¢yt¡1 ¡ ¹St¡1;Vt¡1) + ²t; ²t » N(0; ¾Vt). The parameter V

indicates the variance state, with V =1 corresponding to the high variance state and V =2

to the low variance state, so that ¹11 indicates the mean in the high mean, high variance

state. pkij indicates the probability of switching from state i to state j for the parameter k.

The bottom panel reports the p-value for test of the restriction that the mean growth rate

in recessions is equal across high and low variance states, and that the mean growth rate

during expansions is equal across high and low variance states.

12

To answer this question we estimate the following model:

¢yt = ¹St + Á(¢yt¡1 ¡ ¹St¡1) + ²t; ²t » N(0; ¾) (9')

which is simply a restricted version of Equation 9, with all terms de¯ned as in that

equation. We refer to Equation 9' as the `Hamilton' model because, following Hamil-

ton (1989), the variance is modeled as constant. We plot the smoothed probability of

the low mean state from our augmented model in the middle panel of Figure 3, and

the analogous probability for the Hamilton model in the bottom panel.

The most striking di®erence between these two plots is in the behavior of the

smoothed probabilities during the recessionary period of the early 1990's. While the

Hamilton model virtually misses this recession, the augmented model easily captures

it. This exercise illustrates the importance of accounting for the volatility break in

this type of state-space characterization of business cycle °uctuations: When the

model is not augmented to allow for the dramatic change in the variance, the signal

from the 1990-91 recession is simply too weak to register this period as a recession.7

Turning back to our point estimates, we see that the average rates of output

growth in both expansions and recessions are lower in absolute value for the years

since the early 1980s. Can one conclude from this that the business cycle has been

muted, or `tamed' in recent years, as suggested, for example, in Kim and Nelson

(1999)?

One way to test this hypothesis is to impose the restriction that ¹11=¹12 and

¹21=¹22. To the extent that the mean rates of output growth rates during recessions

and expansions are a cyclical characteristic of output growth, a rejection of this

restriction would suggest some fundamental change in the business cycle. As shown

in the bottom panel of Table 3, however, we cannot reject this restriction. Though not

reported in the table, we also test each element of the above restriction separately. In

7See Estrella and Mishkin (1998), Kim and Murray (1998) and Hamilton and Perez Quiros (1996)for examples from the literature in which state space models that do not account for the volatilitydecline fail to identify the recession of the early 1990s.

13

Figure 3: Markov-Switching Model of U.S. Real GDP Growth: 1953:2 to 1999:2

1953.2 1956.2 1959.2 1962.2 1965.2 1968.2 1971.2 1974.2 1977.2 1980.2 1983.2 1986.2 1989.2 1992.2 1995.2 1998.2-4

-2

0

2

4

0

0.2

0.4

0.6

0.8

1

1953.2 1956.2 1959.2 1962.2 1965.2 1968.2 1971.2 1974.2 1977.2 1980.2 1983.2 1986.2 1989.2 1992.2 1995.2 1998.2-4

-2

0

2

4

0

0.2

0.4

0.6

0.8

1

1953.2 1956.2 1959.2 1962.2 1965.2 1968.2 1971.2 1974.2 1977.2 1980.2 1983.2 1986.2 1989.2 1992.2 1995.2 1998.2-4

-2

0

2

4

0

0.2

0.4

0.6

0.8

1

Probability of Low Mean

Probability of Low Variance

Hamilton Model

Augmented Model

Probability of Low Mean

GDP Growth

Augmented Model

14

each case we cannot reject the null hypothesis. Thus, while we ¯nd strong support for

the hypothesis that quarter-to-quarter °uctuations in GDP growth have been muted,

we do not ¯nd statistical support for the notion that expansions are now less robust

and recessions less severe.8 At the very least, it would seem that it is too soon to tell.

The implications of the volatility break extend well beyond those for state-space

models. The decline should be accounted for in empirical macroeconomic techniques

ranging from model calibration to the estimation of structural vector autoregression

models over periods spanning the break. It also implies that policymakers and eco-

nomic analysts should update their posterior distribution of quarter-to-quarter GDP

growth to re°ect the fact that extreme movements in output are much less likely

to occur today than they were twenty or thirty years ago. For example, during the

period 1953:2 to 1983:4, approximately 30% of quarterly GDP growth rates were in

excess of 1.5%, while for the period beginning in 1984:1, only 3.2% of observations

were as large as 1.5% (in fact, if we used 1984:3 rather than 1984:1, no quarterly

growth rates would exceed 1.5%). On the other end of the distribution, realizations

of output growth below 0% accounted for 22% of the total in the early period, but

for only 4.8% for the period since 1984.

8Up to this point we have not discussed the case in which the variance is constant but the meansin the later period are allowed to be di®erent from those of the early period. The obvious di±cultyhere is de¯ning the `later' and `earlier' periods if they are not de¯ned by the switching variance. Tocorrectly address this issue one needs a test in the spirit of Hansen (1992b) and Hansen (1994). Giventhe computational intensity of such a test we instead undertake the following exercise. We add adummy variable to the augmented model that is de¯ned to be zero before 1984:1 and one thereafter.We then impose two sets of restrictions. The ¯rst is identical to the one described above, in whichthe ¹ and Á parameters are set equal across variance states. We again ¯nd that we cannot rejectthis restriction (p-value=0.48). Next, we relax that restriction and impose one in which ¾1=¾2. Weare easily able to reject this restriction (p-value=0.00). We are therefore reasonably con¯dent thatwe have two di®erent variances over our sample, rather than two di®erent means.

15

3 Sources of the Decline in Output Volatility

3.1 A Look at the Disaggregate Data

We analyze the sources of decreased volatility by decomposing GDP growth into

expenditures on goods, services and structures. For each of these product types, we

¯t AR models to both the growth contribution and the growth rate. We then test

for breaks in the residual variance and AR coe±cients. Throughout our analysis, we

compute growth contributions as the product of the share of nominal GDP accounted

for by a particular component in period t-1 and the real growth rate of that component

in period t.9 We examine contributions because a break in the growth contribution of

an individual component signals a potentially causal role for that component in the

aggregate volatility decline. Further, we examine rates because a break in the growth

rate of a particular component indicates whether the break in the growth contribution

is emanating from increased stability within that sector, or whether there has instead

been a change in the share of output accounted for by that sector.

The results of these product-type break tests are reported in Panel A of Table 4.

Results for the growth contributions are given in the second through ¯fth columns,

while the results for the growth rates are presented in the last four columns.10 We

identify a break in the growth contribution of goods in 1984:1, and in the growth

contribution of structures in 1990:4. Looking now within sectors, we ¯nd that the

growth rates of both the goods and services sectors stabilized during our sample.

The volatility decline in goods, however, is dated 1984:1, and thus lines up with the

9The BEA uses a more complicated method to compute the quarterly growth contributions. Weused annual data, however, to compare our method with the BEA's. The correlation between theBEA's growth contributions and those computed using the lagged nominal weights is greater than0.99.10With the exception of the second entry in Panel C of this table, all results refer to break tests

on the residual variance from an optimally chosen AR speci¯cation for the variable listed in the ¯rstcolumn. Results of the tests for breaks in the AR coe±cients are omitted because we do not rejectstability in any of the cases. We also tested for breaks in the covariances between each of theseproduct types and ¯nd signi¯cant breaks only when there is also a signi¯cant break in the varianceof one of the two components. Thus we are not able to extract any independent information fromthe analysis of the covariances.

16

aggregate volatility reduction, while the timing of the increased stability in services,

1967:1, does not.

Since we ¯nd a break in both the growth rate and growth contribution of the

goods sector in 1984, we look further into this sector, putting aside for the moment

the early 1990's break in the growth contribution of structures. Decomposing goods

growth into contributions from growth in durables and nondurables, we again test for

breaks in the residual variance and AR coe±cients of the growth contributions and

growth rates. Panel B of Table 4 reports that both the growth rate and contribution

of durables break in the ¯rst quarter of 1985. The growth contribution of nondurables

breaks in 1990:4, the same date as the structures contribution, but there is no evidence

of increased stability within the nondurables sector itself.

This investigation reveals that the volatility of durable goods, total goods and

aggregate GDP all declined in the early to mid 1980's, a fact that suggests that the

reduction in aggregate volatility emanated from changes within the durable goods

sector. We also ¯nd, however, breaks in the growth contributions of structures and

nondurable goods in the early 1990's. In order to gauge the importance of these two

¯ndings for the behavior of aggregate volatility, we undertake two simple experiments.

First, to assess the role of the decline in durables volatility in explaining the

aggregate reduction, we generate a new series for durable goods in which the variance

throughout our entire sample is assumed to be that of the pre-1985 period. We then

use this new series to generate an aggregate output series (labeled GDPexp1), and

test this series for the presence of a structural break. The result, reported in Panel

D of Table 4, indicates that the output series generated under the counterfactual

assumption of constant volatility in the durable goods sector does not experience a

volatility break during our sample. Thus, the magnitude of the decline in durables

volatility alone is su±cient to account for the break in the volatility of aggregate

output.11

11This experiment is not strictly correct in that we should allow the weights to change each periodas the growth rate of durables changes. Our omission of this portion of the exercise disadvantagesour hypothesis that the reduction in the volatility of durables alone can account for the reduction

17

The second experiment evaluates the importance of the breaks in the growth

contribution of durables and structures. First, it is useful to note that a break in

the growth contribution of a particular component, in the absence of a corresponding

break in its growth rate, indicates a change in the nominal share of that sector. Indeed,

the nominal share of each of these sectors has changed over the course of our sample.

Using 1990:4 to split the sample (the date of the breaks in the growth contributions

of structures and nondurables), we ¯nd that nondurables accounted for, on average,

27 percent of output pre-1990:4 and only 21 percent post-1990:4. Likewise, structures

declined from approximately 11 percent of output in the early period to 9 percent

in the later period. Both of these declines, along with a slight decline in the share

of durables output, have translated into an increase in the share of nominal output

accounted for by the more stable service sector from 44 percent to 54 percent.

Our interest is in assessing the extent to which these changes in the nominal shares

of each sector have a®ected aggregate output volatility. To do this, we hold each

sector's share constant at its sample wide average, thereby not allowing the shares

of structures and nondurables to decline in the way that they actually did. Under

this assumption we generate a new output series (GDPexp2) and test this series for

a break. Panel D of Table 4 reports that this series does experience a reduction in

volatility in the second quarter of 1984. From this we conclude that changes in the

composition of output, and the corresponding breaks in the growth contributions of

structures and nondurables, are not su±cient to explain the reduction in aggregate

volatility in the early 1980's.

Since our analysis points to a causal role for changes within the durable goods

sector in stabilizing the aggregate economy, we consider the following decomposition

in the volatility of GDP.

18

Table 4: Structural Break Tests: Disaggregate U.S. real GDP Growth - 1953:2 to1999:2

Growth Contributions Growth RatesComponent Date Exp Ave Sup Date Exp Ave Sup

Panel AGoods 1984.1 0.00 0.00 0.00 1984:1 0.01 0.00 0.01Services none 0.40 0.13 0.08 1967:1 0.09 0.02 0.00Structures 1990:4 0.03 0.01 0.04 none 0.41 0.25 0.26

Panel BDurables 1985:1 0.01 0.00 0.00 1985:1 0.00 0.00 0.00Nondurables 1990:4 0.07 0.04 0.07 none 0.54 0.36 0.34

Panel CGDPexp1 none 0.73 0.37 0.29GDPexp2 1984:2 0.01 0.00 0.00

Panel DDurable Final Sales none 0.55 0.62 0.72Absolute Value of (¢I=Dur) 1984:4 0.00 0.00 0.00

Panel EGDB 1984:1 0.00 0.00 0.00FSD none 0.45 0.33 0.30

Note: With the exception of the second entry in Panel D, all results refer to break tests

on the residual variance from an optimally chosen AR speci¯cation for either the growth

contribution or growth rate of the variable listed in the ¯rst column. Results for growth

contributions are reported in the second through ¯fth columns, while growth rate results

are reported in the last four columns. Results of the tests for breaks in the AR coe±cients

are omitted because we do not reject stability in any of the cases. `Sup', `Exp' and `Ave'

refer to the supremum, exponential and average test statistics described in Equations 5, 6,

and 7, respectively. Reported numbers are p-values of indicated test statistic. Estimated

break dates are reported only when at least one test indicates signi¯cance at the 5 percent

level. In Panel C, GDPexp1 is simulated output under the assumption of no reduction in

the volatility of the durable goods sector after 1984:1. GDPexp2 is generated under the

assumption of constant nominal shares of the components of GDP growth. In Panel E,

GDB is gross domestic purchases and FSD is ¯nal shares to domestic purchasers.

19

of the growth rate of durable goods production:12

¢durt = ¢salt(sal

dur)t¡1 +¢iit(

ii

dur)t¡1 (10')

where ¢sal is the growth rate of real sales of durable goods and ¢ii is the growth rate

of real inventory investment. We test for breaks in ¢sal and iidur, but do not analyze

¢iit since in addition to the obvious computational di±culties associated with its

construction, it lacks a clear economic interpretation.

The structural break test results are reported in Panel D of Table 4. The ¯rst line

in Panel D shows that there is no evidence of a change in the volatility of sales growth

over our sample period. That we ¯nd no reduction in the volatility of durables sales,

but strong evidence of a break in durables production, is unsurprising when we view

plots of these two series. Figure 4 shows that for the period prior to the early 1980's,

production is considerably more variable than sales, while for the period since, the

variability of production dampens to a level apparently comparable to that of sales.

In contrast to the constant volatility of sales growth over our sample, the second

line of Panel C reports strong evidence of a break in j¢invdurj in the third quarter of 1984,

a date that corresponds closely to that found for aggregate output. Note that we test

the absolute value of iidur

because we are interested in determining whether inventory

movements, either positive or negative, have become a smaller fraction of durables

production. Our results indicate that this is precisely what has happened{for the

period since the early 1980's, the average share of this extremely volatile component

of durables output is approximately half its previous value.

3.2 Discussion

Our decomposition of U.S. real GDP growth sheds light on the economic sources of

the volatility decline. For example, while one might be tempted to attribute the

12Though our interest is in the variance of durables growth rather than its level, we simplify theanalysis by testing for breaks in the terms of Equation 10' since the terms of the expression for thevariance are only more complicated functions of the former.

20

Figure 4: Production and Sales of Durable Goods - 1953:2 to 1999:2

1953.2 1957.2 1961.2 1965.2 1969.2 1973.2 1977.2 1981.2 1985.2 1989.2 1993.2 1997.2-15

-10

-5

0

5

10

15

Production Sales

increased stability to a shift towards services and away from goods production, our

analysis indicates that such compositional shifts have been of little importance for

stabilizing real GDP growth. Further, since the aggregate volatility drop stems from

a reduction in volatility within the durable goods sector itself, its source is clearly not

a shift in the composition of output across broad sectors of the economy.13 Further,

since the aggregate volatility drop stems from a reduction in volatility within the

goods sector itself, its source is clearly not a shift in the composition of output across

broad sectors of the economy.

Looking in another direction, there is a growing literature documenting changes in

the conduct of U.S. monetary policy before and after 1979 and discussing the extent

to which such changes are likely to have rendered policy a more stabilizing in°uence in

the later period (see for example, Clarida, Gal¶³ and Gertler (1998)). While the timing

of this explanation lines up closely enough with the volatility decline documented in

this paper, any argument for an important role for policy would have to explain why

13Filardo (1997) points out that while employment shares in manufacturing have fallen dramati-cally, productivity increases in that sector have o®set the e®ects of declining employment on output.

21

we observe a decline only in durables volatility, with no corresponding decline in the

volatility of any other sector of the economy. Further, even if one can argue that

the interest sensitive durable goods sector might be the most responsive to policy,

it is di±cult to imagine a model in which policy has e®ects on the production of

durable goods, but leaves the process for sales una®ected. Though an important role

for policy cannot be ruled out by our decomposition, neither is it easy to construct a

policy related story consistent with the facts presented here.

One might also hypothesize that trade patterns shifted in such a way that the

U.S. began importing, rather than producing, goods from relatively volatile sectors.

To examine this hypothesis, we conducted structural break tests on gross domestic

purchases of goods and services (denoted GDB), which is essentially gross domestic

production with imports added back in and exports subtracted out. The ¯rst row

of panel D in Table 4 shows that the volatility decline is also a feature of domestic

purchases, suggesting that changes in trade patterns are not an important factor in

the volatility decline.

None of the above scenarios explains why we observe a dramatic decline in durables

production volatility, but no decline in sales volatility. Nor do they explain why a

sharp reduction in the share of durables output accounted for by inventories occurred

around the same time as the aggregate volatility decline. These stylized facts sug-

gest an important role for durables inventories in explaining the aggregate volatility

decline.

One hypothesis is that changes in inventory management, such as the use of `just-

in-time' techniques, have brought about the reduced share of durables inventory and

the associated stability of aggregate output.14 While we do not test this hypothesis

here, we highlight a number of patterns in the data that are suggestive of improved

14Discussion of just-in-time techniques can be found Freeland (1991), Morgan (1991), Bechterand Stanley (1992), Norris et al. (1994) Allen (1995), Filardo (1995), Flood and Lowe (1995),Hirsch (1996), Ramey and West (1997). Not surprisingly, there is also much discussion of improvedinventory management in trade journals and the business press, examples include Business Week(April 4, 1994), Supply Management (July 16, 1998), Barron's (May 25, 1998), and Purchasing(September 1, 1998).

22

Figure 5: Inventory to Sales Ratios: 1953:2 to 1997:2

1959.01 1962.01 1965.01 1968.01 1971.01 1974.01 1977.01 1980.01 1983.01 1986.01 1989.01 1992.01 1995.01

1.6

1.8

2

2.2

2.4

2.6

1

1.1

1.2

1.3

I/S-Durables I/S-Nondurables

I/S-Durables I/S-Nondurables

inventory management.

For example, inventory-to-sales ratios in durables, as shown in Figure 5, have

been declining almost steadily since the early 1980's. Interestingly, no such change

in trend is evident in the nondurable goods sector. Perhaps more striking is evidence

from a monthly survey conducted by the National Association of Purchasing Man-

agers, which prompts respondents to report how many days in advance they order

production materials. Though not reported here, we ¯nd a statistically signi¯cant

increase in the early 1980's in the proportion of respondents ordering their production

material hand-to-mouth (15 days or less), as well as in those ordering 30 days or less.

In addition, the average lead time across all respondents for the period from January

1961 to December 1983 was seventy-two days, while it was only forty-nine days for

the 1984-98 period.

Though our discussion of inventories has concentrated on the durables goods sec-

tor, we undertake one ¯nal test to illustrate the importance of changes in inventory

behavior for aggregate GDP volatility. In particular, we subject ¯nal sales to domestic

purchasers (FSD), de¯ned as GDB minus imported as well as domestically produced

23

inventories, to a structural break test of the form conducted throughout this paper.

In the bottom row of Table 4, we see that FSD does not have a volatility break. Thus,

the volatility break is eliminated from domestic purchases by subtracting purchases

of inventories from total purchases. In other words, there has been no change in the

volatility of non-inventory purchases. Clearly some aspect of inventory investment in

the U.S. has changed in such a way as to have had a strong e®ect of the volatility of

U.S. output °uctuations.

A complete characterization of the mechanism by which inventories are an impor-

tant factor in producing the lower volatility surely requires the use of disaggregate

data. Industry-level data should allow one to distinguish a scenario in which the

source of stability is improved inventory management techniques from one in which

the declining share of inventory investment re°ects a shift toward less inventory-

intensive industries. Similarly, one could investigate whether shifts in the composition

of goods by stage of processing has e®ected a reduction in the fraction of inventories

held in the aggregate.

4 Conclusions

This paper identi¯es a structural decline in the volatility of quarter-to-quarter °uctu-

ations in U.S. GDP growth in the early 1980's. We trace the source of the reduction

in aggregate volatility to a decline in the volatility of durables output and provide

evidence that the timing of the drop in durables volatility corresponds to a reduction

in the share of durables output accounted for by inventory investment.

This break in volatility has important implications for widely used theoretical

and empirical techniques, examples of which include the estimation of state-space

models of business cycle °uctuations, model calibration exercises and the estimation of

structural vector autoregression models over periods spanning the break. In addition,

since the break implies that we are now much less likely to see extreme movements

in GDP growth, it a®ects the interpretation policymakers place on particular growth

24

rate realizations.

Finally, while the evidence in this paper points only to a change in quarter-to-

quarter rather than cyclical °uctuations in output, one interesting aspect of the cy-

cle that we have not explicitly considered is that documented in Sichel (1994). In

particular, Sichel ¯nds that there are typically three phases of a business cycle; re-

cessions, high growth recoveries and moderate growth periods following recoveries.

He attributes this three phase pattern for the post-war period to swings in inventory

investment. Moreover, he points out that there was not a high growth recovery fol-

lowing the recession of 1990-91 and attributes this to a combination of both weak

inventory investment and weak ¯nal sales. The question of why sales growth was

so weak following this recession not withstanding, one would like to know whether

inventory investment tracked sales more closely during the early years of this recov-

ery than during other recoveries, thus rendering the recovery more sluggish than it

otherwise might have been. The diminished role of inventories documented here may

serve to substantially weaken the high growth recovery phase.

25

References

Allen, Donald S. \Changes in Inventory Management and the Business Cycle." Re-

view Federal Reserve Bank of St. Louis, July/August 1995, 74(4), pp. 17-26.

\Do Inventories Moderate Fluctuations in Output." Review Federal Reserve

Bank of St. Louis, July-August 1997, 79(4), pp. 39-50.

Andrews, Donald. \Test for Parameter Instability and Structural Change with Un-

known Change Point." Econometrica, 61(4), July 1993, pp. 821-56.

Andrews, D. and Ploberger, W. \Optimal Tests when a Nuisance Parameter is Present

only under the Alternative." Econometrica, November 1994, 62(6),pp. 1383-1414.

Bechter, Dan M. and Stanley, Stephen. \Evidence of Improved Inventory Control."

Economic Review, Federal Reserve Bank of Richmond, Jan/Feb 1992, 78(1), pp.

3-12.

Boldin, Michael D. "Dating Turning Points in the Business Cycle" Journal-of-Business,

January 1994, 67(1), pp. 97-131.

Clarida, Richard; Gal¶³, Jordi and Gertler, Mark. \Monetary Policy Rules and Macro-

economic Stability: Evidence and Some Theory."' NBER Working Paper No.

W6442, March 1998.

Davidian, M. and Carroll, R.J. \Variance Function Estimation." Journal of the Amer-

ican Statistical Association, December 1987, 82(400), pp. 1079 - 1091.

DeLong, J. Bradford and Summers, Lawrence. \The Changing Cyclical Variability

of Economic Activity in the United States," in R.J. Gordon, ed., The American

Business Cycle, Chicago: University of Chiacgo Press, 1986, pp 679-734.

Diebold, Francis X, and Rudebusch, Glenn. \Have Postwar Economic Fluctuations

Been Stabilized?" American Economic Review, September 1992, 82(4), pp. 993-

1005.

26

_\Measuring Business Cycles: A Modern Perspective". Review of Economics

and Statistics, February 1996, 78(1), pp. 67-77.

Durland, J. Michael, and McCurdy, Thomas H. \Duration Dependent Transitions in a

Markov Model of US GNP Growth." Journal of Business and Economic Statistics,

July 1994, 12(3), pp. 279-88.

Epstein, Gene. \Looking for Proof in All the Wrong Places, Economists Miss the

Point on PCs and Output." Barron's, May 25, 1998.

Estrella, Arturo and Mishkin, Frederic S. \Predicting U.S. Recessions: Financial

Variables as Leading Indicators." Review of Economics and Statistics, February

1998, 80(1), pp 45-61.

Filardo, A.J. \Business Cycle Phases and their Transitional Dynamics." Journal of

Business and Economic Statistics, July 1994, 12(3), pp. 299-308.

_\Recent Evidence on the Muted Inventory Cycle." Economic Review Federal

Reserve Bank of Kansas City, 2nd Quarter 1995, 80(2), pp. 27-43.

_\Cyclical Implications of the Declining Manufacturing Employment Share."

Economic Review Federal Reserve Bank of Kansas City, 2nd Quarter 1997, 82(2),

pp. 63-87.

Flood, Darren, and Lowe, Philip. \Inventories and the Business Cycle." The Eco-

nomic Record, March 1995, 71(212), pp. 27-39.

Freeland, James R. \A Survey of Just-In-Time Purchasing Practices in the United

States," Production and Inventory Management Journal, Second Quarter, 1991.

Gleckman, Howard and Schiller, Zachary. \A Tonic for the Business Cycle." Business

Week, April 4, 1994.

Hamilton, James D. \A New Approach to the Economic Analysis of Nonstationary

Time Series and the Business Cycle." Econometrica, March 1989, 57(2), pp. 357-

84.

27

and Perez Quiros, Gabriel. \What do the Leading Indicators Lead?" Journal

of Business, January 1996, 69(1), pp. 27-49.

Hansen, Bruce E. \Testing for Parameter Instability in Linear Models." Journal of

Policy Modeling, August 1992a, 14(4), pp. 517-33.

_̀`The Likelihood Ratio Test Under Non-standard Conditions: Testing the

Markov Trend Model of GNP." Journal of Applied Econometrics, Suppl. Dec.

1992b, 7(0), pp. S61-82.

_\Erratum: The Likelihood Ratio Test Under Non-standard Conditions: Test-

ing the Markov Switching Model of GNP." Journal-of-Applied-Econometrics, March-

April 1996, 11(2), pp. 195-98.

_\Approximate Asymptotic p-values for Structural Change Tests," Journal of

Business and Economic Statistics, January 1997, 15(1), pp. 60-67.

_\Testing for Structural Change in Conditional Models." forthcoming in The

Journal of Econometrics, March 1999.

Hess, Gregory D. and Iwata, Shigeru."Measuring and Comparing Business-Cycle Fea-

tures" Journal-of-Business-and-Economic-Statistics, October 1997, 15(4), pp. 432-

44.

Hirsch, Albert A. \Has Inventory Management in the US Become More E±cient

and Flexible? A Macroeconomic Perspective" International Journal of Production

Economics, August 1996, 45 (1-3), pp. 37-46.

Kim, C.J. \Predicting Business Cycles Phases with Indexes of Leading and Coincident

Economic Indicators: A Multivariate Regime-Shift Approach"Journal of Economic

Theory and Econometrics, 1996, 2(2), pp. 1-27.

and Murray, Chris. \Permanent and Transitory Components of Recession-

s"University of Washington Discussion Paper, May 1999.

and Nelson, Charles R. \Has the U.S. Economy Become More Stable? A

Bayesian Approach Based on a Markov-Switching Model of the Business Cycle"

28

Review of Economics and Statistics, November 1999, 81(4), pp. 1-10.

Koop, Gary and Potter, Simon, \Nonlinearity, Structural Breaks or Outliers in Eco-

nomic Times Series?" in William A. Barnett., et al, eds., Nonlinear Econometric

Modeling in Time Series Analysis, Cambridge University Press, 2000.

Lam, Pok-Sang. \The Hamilton Model with a General Autoregressive Component."

Journal of Monetary Economics, December 1990, 26(3), pp. 409-432.

Lebergott, Stanley, "Discussion of Romer and Weir Papers." Journal-of-Economic-

History, June 1986, 46(2), pp. 367-71.

McConnell, Margaret and Gabriel Perez Quiros. \Output Fluctuations in the United

States: What Has Changed Since the Early 1980?" Sta® Reports Federal Reserve

Bank of New York, June 1998, (41).

Minahan, Tim. \JIT Moves Up the Supply Chain." Purchasing, September 1, 1998.

Morgan, Donald P. \Will Just-In-Time Inventory Techniques Dampen Recessions?"

Economic Review, Federal Reserve Bank of Kansas City, April 1991, pp. 21-33.

Norris, Daniel M.; Swanson, Robert D. and Chu, Yung-Lin. \Just-In-Time Produc-

tion Systems: A Survey of Managers." Production and Inventory Management

Journal, Second Quarter 1994.

Politis, Dimitris N. and Romano, Joseph P., "The Stationary Bootstrap" Journal-of-

the-American-Statistical-Association, December 1994, 89(428), pp. 1303-13.

Ramey, Valerie A. and West, Kenneth D.. \Inventories." Manuscript, 1997.

Ramey, Garey and Ramey, Valerie A. \Cross-Country Evidence on the Link Between

Volatility and Growth." American Economic Review, 1995, 85(5), pp. 1138-1151.

Romer, Christina D. \Spurious Volatility in Historical Unemployment Data." Journal

of Political Economy, February 1986, 94(1), pp. 1-37.

_\Is the Stabilization of the Postwar Economy a Figment of the Data?" Amer-

ican Economic Review, June 1986, 76(3), pp. 314-334.

29

_\The Prewar Business Cycle Reconsidered: New Estimates of Gross National

Product, 1869-1908." Journal of Political Economy, February 1989, 97(1), pp.

1-37.

\Remeasuring Business Cycles." The Journal of Economic History, Septem-

ber 1994, 54(3), pp. 578-609.

Shapiro, Matthew D. \The Stabilization of the U.S. Economy: Evidence from the

Stock Market." American Economic Review, December 1988, 78(5), pp. 1067 -

1079.

Sichel, Daniel E. \Inventories and the Three Phases of the Business Cycle." Journal

of Business and Economic Statistics, July 1994, 12(3), pp. 269 - 277.

Skedd, Alex. \Time is of the Essence." Supply Management, July 16, 1998.

Watson, Mark W. \Business-Cycle Durations and Postwar Stabilization of the U.S.

Economy." American Economic Review, March 1994, 84(1), pp. 24-46.

West, Kenneth D. \A Comparison of the Behavior of U.S. and Japanese Inventories."

International Journal of Production Economics , 1992, 26, pp. 115-222.

30